Kimi K3.1 Pro Goes Open Weight: Benchmarks and Pricing
Share: https://www.ellixorlabs.com/blog/kimi-k3-1-pro-open-weights-benchmarks-pricing (copy URL for iMessage or social)
Published (display in IST)
Updated (IST)

Moonshot AI has released the full weights of Kimi K3.1 Pro, a 2.8 trillion parameter model it says rivals GPT-6 Astra and Claude. Here is how it scores, what it costs and what it means for builders.
Cover image: Moonshot AI, Kimi K3.1 Pro announcement. View source.
A 2.8 trillion parameter AI model that its maker says trades blows with the best from OpenAI and Anthropic is now free to download. On 4 November 2026, Chinese AI lab Moonshot AI published the full weights of Kimi K3.1 Pro on Hugging Face, keeping the promise it made when the model went live in its apps and API on 23 October.
In short, Kimi K3.1 Pro now ranks as the top open-weight model on at least one independent index, it costs $2 per million input tokens through Moonshot's API, and its licence lets most businesses use and modify it for free. The catch is size: you will not be running it on a laptop.
That matters now because the gap between models you rent and models you can own keeps narrowing. For founders building AI products, every open release at this level is fresh leverage on price, privacy and vendor lock-in.
What Moonshot released, and when
Moonshot unveiled K3.1 Pro on 23 October as the first model in its new K3.1 series, describing it in its tech blog as its "most capable 3T-class open-weight model to date". It is a post-trained version of Kimi K3, built with what the company calls substantially larger reinforcement learning compute across hundreds of task domains.
The headline specifications, from the official model card:
The weights landed on schedule. ArtificialWatch, which tracks model releases, counted 96 files totalling about 1.56 TB in the public, ungated repository. Moonshot marked the moment on X with a short post linking to the release.
We'll keep advancing open intelligence.
Posted by @Kimi_Moonshot on X, 4 November 2026.
Benchmarks: how it stacks up against GPT-6 Astra and Claude
Most of the numbers below come from Moonshot's own model card, so treat them as self-reported. Moonshot ran its model at maximum reasoning effort and took most rival scores from official leaderboards and system cards. Where Kimi leads:
Where it trails:
An outside view is more measured. On ArtificialWatch's Bench Index, which takes the median percentile across public leaderboards, K3.1 Pro scores 69.9 and ranks 11th of 44 models, open and closed. That makes it the highest ranked model with downloadable weights, narrowly ahead of DeepSeek V4.1 Flash. On the same site's coding index, though, it sits 21st of 45.
Moonshot also reports a sharp drop in made up facts. On its internal stress test, the claim level hallucination rate fell from 23.4% for Kimi K3 to 8.3%. That is an internal benchmark built to provoke errors, and Moonshot itself says the figure is not an everyday error rate.
Did Moonshot hold anything back?
Yes, deliberately. The company says it stripped cyber attack related data from training for the open release. On its internal exploit benchmark, the released model solves 81.0% of tasks, against 90.5% for an otherwise identical checkpoint trained without those controls. Moonshot says it does not want open weights to push the frontier of offensive hacking capability.
What it means for builders and founders
How much does Kimi K3.1 Pro cost?
Through Moonshot's API, the standard model costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.20, according to Kimi's official pricing page. That is a third cheaper than Kimi K3. Highspeed costs double, at $4 and $20, and batch jobs are billed at 60% of the standard price. The API is compatible with OpenAI and Anthropic formats, so trying it in an existing app can be little more than a configuration change.
Can a small team run it themselves?
Not cheaply. At around 1.56 TB of weights, this is data centre hardware territory. AMD, for example, announced day zero support for running it on eight of its Instinct MI355X accelerators. For most startups the practical routes are Moonshot's API or third party hosts that serve the open weights.
What does the licence allow?
Use, modification, fine tuning and redistribution are free, with two conditions. A company selling model access as a service with more than $20 million in revenue over 12 months needs a separate agreement with Moonshot, and a product with more than 100 million monthly users or $20 million in monthly revenue must display the Kimi K3.1 Pro name prominently. Neither touches a typical Indian SaaS startup or agency.
The bigger point is choice. A model you can download is one you can audit, fine tune on private data and keep running if a vendor changes its terms. If you are weighing where AI fits in your product, our guide on when to bring in AI and automation is a sensible starting point, and the recent wave of AI coding tool acquisitions shows how quickly the closed side of the market is consolidating.
What to watch next
Independent testing is the next step. Self-reported benchmarks rarely survive contact with real workloads unchanged, so watch the public leaderboards and third party evaluations over the coming weeks. Watch, too, how quickly hosting providers cut prices now that anyone can serve the weights, and whether Moonshot publishes the Highspeed checkpoint.
If you want to test an open model like this inside your own product or workflows without betting the roadmap on it, the Ellixor Labs team can help you scope a pilot.
Frequently asked questions
Is Kimi K3.1 Pro free to use?
Yes for most businesses. The weights are free to download, modify, fine tune and redistribute. Companies selling model access with over $20 million in revenue over 12 months need a separate agreement with Moonshot, and very large products must credit the model by name.
How much does the Kimi K3.1 Pro API cost?
Moonshot charges $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens for the standard model. The faster Highspeed version costs $4 and $20, and batch jobs cost 60% of the standard price.
Can I run Kimi K3.1 Pro on my own computer?
Not on a normal computer. The 2.8 trillion parameter model ships as about 1.56 TB of weights and needs multi GPU data centre hardware, such as a server with eight AMD MI355X accelerators. Most teams use Moonshot's API or a third party host.
FAQ
Is Kimi K3.1 Pro free to use?
Yes for most businesses. The weights are free to download, modify, fine tune and redistribute. Companies selling model access with over $20 million in revenue over 12 months need a separate agreement with Moonshot, and very large products must credit the model by name.
How much does the Kimi K3.1 Pro API cost?
Moonshot charges $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens for the standard model. The faster Highspeed version costs $4 and $20, and batch jobs cost 60% of the standard price.
Can I run Kimi K3.1 Pro on my own computer?
Not on a normal computer. The 2.8 trillion parameter model ships as about 1.56 TB of weights and needs multi GPU data centre hardware, such as a server with eight AMD MI355X accelerators. Most teams use Moonshot's API or a third party host.
Talk to Ellixor Labs about this topic.
Source: Moonshot AI on Hugging Face